How to judge a stranger in my dataset?

[Tomhouxin]

. I want to achieve real_time_face_recognition py, it mistook a stranger people in the dataset, it should be the most similar as the stranger in the dataset, then how should I process to determine the stranger unknowner? Thank you very much

[Shahnawazgrewal]

Can you please rephrase your question. @Tomhouxin

[Tomhouxin]

When I was doing face recognition, I found that I could only recognize faces in data sets. The problem is when I identify a face that is not in the data set. It actually discriminates me into the most similar human face in a dataset, not a stranger. @Shahnawazgrewal

[Shahnawazgrewal]

for previously unseen identities, you need to extract embedding and then a train classifier to recognize. @Tomhouxin

[Tomhouxin]

According to my understanding, I just need to decide that people I don't know are strangers, and I don't need to put them into the classifier. @Shahnawazgrewal

[jax79sg]

Maybe find out the distribution of confidence levels of different classes for unknown faces. Then from there, see if can come up with a heuristic to determine if its unknown.

[Tomhouxin]

You have a good idea, but people you know and people you don't know are very similar in terms of SVM classification results. @jax79sg

[RedBugYO]

@Tomhouxin How about create a category called as "unknown" and put a lots of face image which you don't have to classify in it then re-train the classifier? I think it may make the classifier better for usage.

[Tomhouxin]

This problem is solved by setting a threshold for the results of the classifier. At the very beginning, I saw that the results of the classification were very close, and there was no threshold. Now when I train the classifier, I add about 100 faces to each person. The result of the classification is obviously different, and then you can set the threshold.（这个问题解决了，把分类器出来的结果设一个阈值。最开始我看分类的结果都很接近，无法设阈值。现在我训练分类器的时候每人加到100张人脸左右。分类出来的结果就很明显有差距了，这时候就可以设阈值） @RedBugYO

[niu-niu]

@Tomhouxin try other classifier, if you find the classification results are very close to each other.

[jiangzidong]

hi @Tomhouxin What does the "threshold" mean in the SVM classifier？ 1） check the probability by model.predict_proba(emb), if the max one is below a threshold, then it is a stranger? or 2) add a new class called "unknown" in to the classifier?

[DawnHH]

@Tomhouxin 所以你还是增加了一个新类叫stranger吗？和embedding预测的类别相似度小于阈值的都被分类为这个类别里？

[rain2008204]

@DawnHH only add a threshold

[Tomhouxin]

@DawnHH yes

[menon92]

Hello @Tomhouxin , Did you solve your problem ?

[Tomhouxin]

@menon92 , yes !

[menon92]

Thanks @Tomhouxin Could you please help me that how can I do this. I have the same problem. I try to calculate `l2 distance' of two image. But it's not working well

Thanks

[Tomhouxin]

sorry，I'm not sure how good it is to calculate the distance. I'm using the classifier (SVM), but I'm also going to try to calculate the distance directly

[menon92]

@Tomhouxin Do you just use classifier (SVM) prediction with a fix threshold to judge a stranger ?

[Tomhouxin]

@menon92 ,yes ,you are right

[menon92]

@Tomhouxin But it's not working for me. what could be the cause of it . However my image quality is not very good

[Tomhouxin]

I don't know exactly what your question is, can't you find the appropriate threshold?

[menon92]

@Tomhouxin Yes, I can't find the appropriate threshold for classifying the unknown images.

[Tomhouxin]

Well, please read the comments carefully. They say how to solve it.

[ZhangYuef]

hi @Tomhouxin What does the "threshold" mean in the SVM classifier？ 1） check the probability by model.predict_proba(emb), if the max one is below a threshold, then it is a stranger? or 2) add a new class called "unknown" in to the classifier?

I think their method is as the same as the first one you mentioned.